Latest issueSunday 8 November 2026 · Issue 10 · 3 min read
AI is building Anthropic's AI twice as fast; Carta prices multiplayer coding at $0.15 per million tokens
The one thing to know
The Verge · yesterday
METR's Anthropic assessment
Takeaway: around twelve times as productive
METR, an AI-safety evaluator, spent eleven weeks inside Anthropic and reported Thursday that the lab's models are improving about twice as fast as a year ago — roughly two years of historical progress in one — with most of that acceleration attributed to AI assisting researchers rather than more training compute. Employees using Anthropic's strongest unreleased internal model reported being around twelve times as productive with it. Three finished models, codenamed Model 2, 3 and 4, have not been released.
The capability assumptions baked into your 2027 roadmap are what is at risk here: if frontier progress is compounding at this rate, features you deferred may arrive as platform capabilities first, and the political heat now on the biggest lab will shape what you can build on its models. Anthropic redacted a majority of the evidence METR found most informative from the public version of its roughly 100-page report; cofounder Sam McCandlish said revealing what drives the speedup is commercially sensitive and hands rivals a roadmap to go faster.
New ground broken
NYT Cooking AI assistant — An AI helper inside the recipe page for iOS subscribers: it rewrites a recipe around what you actually have — swaps, scaling, vegan conversions — files the result under a separate "Your Version" tab, and leaves the original recipe untouched. It was trained on the app's reader comments. This is the pattern to copy for adding an agent to any content product: generate the user's variant, never touch the source of truth, refuse the bad ideas. It declined minced beef in a trifle, but a mid-bake question returned the AI version's six tablespoons of butter instead of the original three-quarter cup — the variant can't be saved separately. The Verge
“works well as a sous chef, but it's not the head chef”
Agent features and products that impact product management
Muse specialized models — Specialized Muse models, trained with Meta's Muse Wonder series, are rolling out now to Muse and Meta AI: Active Memory remembers what matters across conversations and drops what is stale, and Active Focus keeps working toward a goal when something unexpected goes wrong. They arrive alongside last week's "spaces" — mini-apps Muse builds from a description. This sets the consumer-assistant bar at persistence and recovery, not chat quality — an assistant feature that forgets or gives up mid-task will feel broken next to it. Meta reports a 2.4x reduction in forgetting critical information, and said Muse hit 10 million US and Canada users by October's earnings call. Meta AI Research
Evergreen by Carta — A cloud coding platform where teammates and agents work in shared sessions and agents push their own pull requests; work survives a closed laptop. Carta says weekly merged PRs grew from about 1,300 in January to nearly 20,000, and more than 80% of the company now uses it. The bottleneck is shifting from model capability to workflow and controls — shared sessions, audit logs, spending caps — so compare this against per-seat IDE licences, not just against models. Model costs pass through at provider prices plus $0.15 per million tokens from Evergreen, with storage at $0.35 per GB-month, no seat fees and a 7-day trial; it runs Claude, Codex, Grok and Muse. Carta
Market and platform news
ChatGPT connector routing flaw — A developer reported that with two connectors pointed at the same MCP server, ChatGPT sent a follow-up tool call through the second connector without telling the user, carrying a one-time confirmation code issued by the first connector's widget. If you ship an MCP connector, the boundary users care about is the app they authorized, not a shared server URL — test for silent cross-app routing before launch. The server rejected the call because the code was bound to the first connector's OAuth client. OpenAI Developer Community
Ocient — The Chicago data-analytics company raised a $100 million Series C co-led by Boka Capital and OCA Ventures, earmarked for growing its agentic AI analytics business. Analytics infrastructure built for agents querying data is now a funded category, so expect it on procurement shortlists next to your dashboard tools. SiliconANGLE via Techmeme
The big AI stories today
Public First — An Anthropic-backed AI-safety super PAC spent $63M+ in the midterms and 10 of its 17 candidates won, making AI compliance an earlier roadmap item. Bloomberg via Techmeme
Nvidia espionage memo — A US intelligence memo says an executive was recruited by China's Ministry of State Security to spy on Jensen Huang; chip-policy noise keeps GPU supply planning volatile. TechCrunch
AI CEO-spoofing indictment — The DOJ indicted three men who used AI voice and image spoofs of executives and lawyers to steal about $129M; executive impersonation is now a standard fraud playbook. The Record
One takeaway
The Verge · yesterday
Stop pricing your roadmap around what models can't do yet
Between an outside evaluator finding frontier research compounding faster than expected and Meta shipping persistence behaviours to mass-market users, the safe planning assumption this quarter is that capability curves move faster than annual plans allow. Pick the one memory or recovery behaviour your assistant feature currently lacks and design for it getting cheap next year.
One thing to try: list every roadmap item whose case rests on "models can't do that yet" and re-test the top three against current models this month — several will have quietly become buildable.